Sub-pixel accuracy in holography optimize method

[amarosanz]

Hi all,

The optimize method in the holography module takes the Hologram attribute target as its main argument. According to the documentation, the target pattern must be defined as a NumPy/CuPy array.

As far as I understand, this implies that each target light spot must be located at a discrete pixel position, meaning that sub-pixel target locations are not directly supported. For example, if I wanted to place a light point at a position such as (1.5, 2.5), I would need to increase the resolution (e.g. by a factor of two) so that this point coincides with a pixel rather than lying between pixels.

Could you please clarify if this interpretation is correct?
If so, is increasing the resolution the recommended approach for handling sub-pixel target positions, or is there an alternative method (e.g. interpolation or a different target representation) that allows for sub-pixel accuracy?

I ask this also in light of other hologram computation approaches (for example wavelet-based methods) that directly take spot positions as input rather than a pixelated target field.

Thank you for your help,

Amaro

[ichristen]

Hi @amarosanz ,

Your description is correct (though optimize itself does not take in target, instead target is an argument for the Hologram constructor).

Yes, for the discrete Fourier transform (DFT)-based implementation of holography, padding is the solution to improve resolution. There's a helper function get_padded_shape() (link) to help with calculating the needed shape if you want to target some numeric resolution.

Fortunately, we do have other hologram approaches implemented. If you're making a array of spots, you can use CompressedSpotHologram (link). This accepts an array of floating-point coordinates, free from the grid (alongside other stuff too). Keep in mind that this is slow on a CPU alone for large numbers of spots; it will run better on a CUDA-capable GPU.

Best,
Ian

[amarosanz]

Thank you!

The CompressedSpotHologram class might exactly be what I was looking for :)
I have some questions about its principles and usage, but I'll post them in a separate discussion.

Cheers,
Amaro